feat(s53): prompt-cache/memory program — fill the empty Cache tab + cache-stability fixes

[drwhofan2k18-pixel]

Summary

Completes the S53 prompt-cache/memory program: surfaces real provider prompt-cache metrics (filling the empty Cache tab), fixes the cache-stability bugs that suppress the hit rate, and lands the two open issues from the QA pass. The empty Cache tab on v0.11.0 is now fixed end-to-end.

This branch also carries the S53A/S53B work (prefix-break classification, embedder cache, memoryStats(), /api/memory-status + /api/provider-cache).

What this PR delivers

Issues (from the QA/burn pass)

• TurnStore backend divergence: duplicate (conversationId, turnIndex) append throws on SQLite, silently dups on InMemory #9 — TurnStore duplicate-append divergence (b672d7a): both SqliteTurnStore and InMemoryTurnStore now throw a typed DuplicateTurnError on a duplicate (conversationId, turnIndex). Previously SQLite threw a raw ERR_SQLITE_ERROR (leaky) and InMemory silently stored a second row — the contract diverged. Shared compliance suite now asserts it on both backends.
• npm publish leaks ~180 compiled .test.js files + source maps into the package #10 — npm publish leaks test files (641199e): !-negation patterns in files[] cut the package from 1.8 MB → 939 kB (0 compiled .test.js, 0 leaked source maps). .npmignore doesn't apply when files is set, so exclusions live in package.json.

S53 — Cache tab + memory (display + cost visibility)

• Sub-sprint A ✅ — readProviderCacheStats() + /api/provider-cache endpoint (lifetime aggregates, not the 30-min rolling window).
• Sub-sprint B ✅ — ProviderCacheCard + MemoryEffectivenessCard; pricing.ts computes $ saved from inputRatePerToken (cacheRead×0.9 save, cacheWrite×0.25 premium).
• S53B memory ✅ — memoryStats() + /api/memory-status; TrigramEmbedder FIFO cache (256, MEGACOMPACT_EMBED_CACHE).
• CacheTab.tsx now fetches /api/provider-cache alongside the dedup cards — fixes the empty-Cache-tab root cause (it previously read /api/snapshot = dedup stats = zeros).

S53-D — cache-STABILITY (the hit-rate fix) (9430368)

• RECOMPACT_PCT_DELTA 10 → 50 (env-overridable: MEGACOMPACT_RECOMPACT_PCT_DELTA). The v0.8.6 value of 10 rebuilt the KV prefix on every 10% window growth — the root cause of the observed 55–60% hit rate. Replaying the cached trim view up to 50% growth stabilizes the prefix.
• Skip-path fallback (D7): a skipped compaction now falls back to the cached trim replay when trimCache is valid — never to the full transcript (which was a guaranteed cache miss).

S53-C/D5 — MEGA CACHE flare gate (9430368)

• widget.ts gated the dedup flare on cachePct >= 100. After S53-C, cachePct is the bounded 0–100 provider prompt-cache pct, which silently killed the flare. Both flare sites now gate on megaCacheFlarePct (the dedup hit rate that armed the flare, which legitimately exceeds 100%).

Docs

• docs/BRANCH_GAP_ANALYSIS_REFRESH.md — refreshes the 07-29 branch analysis (only feature/promptcache-stats moved, with planning docs only), verifies all claims against current master, and assesses the S53 implementation.

Verification gate (all green)

npm test           → 925 pass / 0 fail across 93 files
npm run lint       → clean (tsc + guardrails + semantic)
regression_check   → clean
npm pack --dry-run → 939 kB (was 1.8 MB), 0 test files leaked

Notes

• S53A/S53B were authored by opencode on this branch; S53-C/D5 + the two issues + the analysis are my additions on top.
• The local implementation diverges slightly from the upstream spec naming (readProviderCacheStats vs spec's readProviderCacheLifetime; pricing.ts was struck in the spec but built anyway — functionally fine).
• Still out of scope (per spec, future sprints): recall-injection relocation (systemPrompt → tail user-role), DB-mirror high-water mark (O(n²) hot path), PLAN_V2 Phases 2–4.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.